Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Highly-Efficient Persistent FIFO Queues

Panagiota Fatourou, Nikos Giachoudis, George Mallis

TL;DR

The paper tackles the challenge of building efficient, durable FIFO queues on non-volatile memory under crash-recovery by reusing and extending state-of-the-art concurrent queue designs. It systematically studies persistence strategies, introduces PerIQ and PerCRQ with a local-persistence optimization, and then presents PerLCRQ, a durable, scalable FIFO built as a linked list of PerCRQ instances. Empirical results show PerLCRQ delivering at least a $2\times$ throughput gain over the best persistent-queue competitors, while also elucidating a clear trade-off between normal execution persistence and recovery cost. The work also provides a practical framework for simulating failures and measuring recovery, and highlights local-persistence as a general technique to reduce persistence overhead in contention-heavy shared variables.

Abstract

In this paper, we study the question whether techniques employed, in a conventional system, by state-of-the-art concurrent algorithms to avoid contended hot spots are still efficient for recoverable computing in settings with Non-Volatile Memory (NVM). We focus on concurrent FIFO queues that have two end-points, head and tail, which are highly contended. We present a persistent FIFO queue implementation that performs a pair of persistence instructions per operation (enqueue or dequeue). The algorithm achieves to perform these instructions on variables of low contention by employing Fetch&Increment and using the state-of-the-art queue implementation by Afek and Morrison (PPoPP'13). These result in performance that is up to 2x faster than state-of-the-art persistent FIFO queue implementations.

Highly-Efficient Persistent FIFO Queues

TL;DR

The paper tackles the challenge of building efficient, durable FIFO queues on non-volatile memory under crash-recovery by reusing and extending state-of-the-art concurrent queue designs. It systematically studies persistence strategies, introduces PerIQ and PerCRQ with a local-persistence optimization, and then presents PerLCRQ, a durable, scalable FIFO built as a linked list of PerCRQ instances. Empirical results show PerLCRQ delivering at least a throughput gain over the best persistent-queue competitors, while also elucidating a clear trade-off between normal execution persistence and recovery cost. The work also provides a practical framework for simulating failures and measuring recovery, and highlights local-persistence as a general technique to reduce persistence overhead in contention-heavy shared variables.

Abstract

In this paper, we study the question whether techniques employed, in a conventional system, by state-of-the-art concurrent algorithms to avoid contended hot spots are still efficient for recoverable computing in settings with Non-Volatile Memory (NVM). We focus on concurrent FIFO queues that have two end-points, head and tail, which are highly contended. We present a persistent FIFO queue implementation that performs a pair of persistence instructions per operation (enqueue or dequeue). The algorithm achieves to perform these instructions on variables of low contention by employing Fetch&Increment and using the state-of-the-art queue implementation by Afek and Morrison (PPoPP'13). These result in performance that is up to 2x faster than state-of-the-art persistent FIFO queue implementations.
Paper Structure (10 sections, 1 theorem, 6 figures, 6 algorithms)

This paper contains 10 sections, 1 theorem, 6 figures, 6 algorithms.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. LCRQ: A Brief Description
  4. Persistent FIFO Queue Algorithm
  5. Persistent IQ
  6. Persistent CRQ
  7. Local Persistence.
  8. Durable Linearizability.
  9. Persistent LCRQ
  10. Performance Evaluation

Key Result

lemma \@thmcounterlemma

Assume that a dequeue $\mathit{deq}$ reads a value $i$ into $\texttt{Head}$ and assume that $\mathit{deq}$ writes back Head to NVM by a failure $c_k$. Then, at recovery time, the following hold: (a) Head will have a value greater than $i$, and (b) Tail will have a value greater than $i$.

Figures (6)

  • Figure 1: Circular array states. On top of the array are the array indexes. Each cell has a value; the subscript corresponds to the index field value.
  • Figure 2: Performance comparison of PerLCRQ with PBQueue and PWFQueue.
  • Figure 3: Cost of persisting Head and Tail in PerLCRQ.
  • Figure 4: Recovery time of PerIQ as the number of operations increases.
  • Figure 5: Recovery time of PerIQ as the queue size increases.
  • ...and 1 more figures

Theorems & Definitions (2)

  • lemma \@thmcounterlemma
  • proof